Source 1 of 7Peer-reviewed journal article. Primary published version of record read in full for this note (article and correction). The published Correction (PNAS 122(34), e2518204122) corrects only Osbert Bastani's institutional affiliation and changes no method, result, figure or claim.
With AI access students did far better on practice; when it was taken away, unguarded users did worse than those who never had it
What it supports
The three-arm design and prompt-only difference; the 48% and 127% practice gains; the 17% unassisted-exam reduction for GPT Base relative to the no-access control; and the mitigation of the negative effect by GPT Tutor's guardrails — each checked against the published methods and results.
Where the evidence stops
One large high school in Turkey, one semester, mathematics, GPT-4-era tools; the exam measures short-term learning, not long-term skill; the effect sizes are one study's findings, not constants. The guardrails "largely mitigated" the harm rather than proving any deployment safe.
Used for
The Turkish high-school field experiment: nearly a thousand students across four 90-minute sessions; a control group with textbooks and notes and two GPT-4 conditions differing only by an unseen system prompt — a plain assistant ("GPT Base") and a guardrailed tutor ("GPT Tutor") given the teacher's solutions and common mistakes, told not to hand over full solutions, to require students to show their work, to give progressive hints and to ask for an explanation. With access, practice grades rose 48 per cent (GPT Base) and 127 per cent (GPT Tutor) over control; on the later closed-book, closed-laptop exam the GPT Base students scored 17 per cent below students who never had access, while the GPT Tutor group was statistically indistinguishable from control.
Source 2 of 7Peer-reviewed journal article. Primary full text read for this note.
AI-assisted people scored higher on reasoning problems but misjudged their own contribution, and so did those working alone
What it supports
The assisted-versus-unassisted scores; that overestimation appeared in both groups so AI did not, in this study, increase the raw calibration error; the similar per-item confidence resolution; the pair-versus-model comparison; and the uninterpreted unaided-estimate figure.
Where the evidence stops
LSAT-style logical-reasoning items and a paid online sample; the studies measure what people produced and what they thought they produced, not what they could do without the tool — that quantity was never measured.
Used for
The reasoning-and-calibration studies: in the cleanly randomised second study, 245 people worked with ChatGPT and 207 without, paid a bonus to estimate their own score accurately; the assisted group scored higher (about 13.5 of 20 against about 9.5) yet estimated about 17; the unassisted group overestimated by a similar margin; both separated right answers from wrong at about the same, modest rate; the human-and-AI pair did not on average beat the model alone, though more than half of participants scored above the model's average; and a recorded estimate of what people thought they could have scored unaided (about 11.6 of 20) was never analysed against actual unaided performance, which was not measured.
Source 3 of 7Peer-reviewed journal article, open access. Directly verified against the published text. Cited as Organization Science 37(2), not the earlier working paper.
Consultants gained on tasks inside AI's competence and lost on a task outside it, without the boundary being visible
What it supports
The 758-consultant preregistered field experiment; the inside-frontier gains (12.2% more tasks, ~25% faster); the nineteen-percentage-point reduction outside the frontier; and the jagged-frontier interpretation.
Where the evidence stops
One professional-services setting and task set; the essay carries no quality-percentage claim from this study, which its argument does not require.
Used for
The claim that assisted consultants completed 12.2 per cent more tasks and worked about 25 per cent faster on work inside the model's competence, but on a complex task deliberately chosen to sit outside it were nineteen percentage points less likely to reach the correct answer than colleagues with no AI — the "jagged frontier," with nothing in the work announcing which side of it a task is on.
Best available evidence
Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, François Candelon & Karim R. Lakhani (2026), "Navigating the Jagged Technological Frontier", Organization Science 37(2), 403–423.
Source 4 of 7Peer-reviewed journal article. Primary published version of record read in full for this note.
When simply asked for the answer, the unguarded assistant was right about half the time, and students used it as a crutch
What it supports
The 51 per cent correct-answer rate; the ~42 per cent logical-error rate; the similar practice-performance coefficients for logical and arithmetic errors; and the authors' crutch/copying interpretation.
Where the evidence stops
The reading is an inference from behavioural and error-rate patterns in one study; no one observed each student to confirm copying directly.
Used for
The error-and-crutch analysis in the same Turkish study: asked directly for the answer, GPT Base returned a correct answer about half the time (51 per cent), with logical errors on roughly four problems in ten; arithmetic and logical errors had similar negative effects on practice performance, even though arithmetic mistakes should have been easier for students to catch — a pattern the authors read as more consistent with copying or disengaged use than with students carefully reading and being misled (logical errors did not spill over to the unassisted exam).
Source 5 of 7Peer-reviewed journal article. Verified; used explicitly as an analogy, once.
Heavier lifetime satnav use went with worse spatial memory when navigating unaided — an analogy only
What it supports
The 50-participant association and the 13-participant three-year follow-up, used as a bounded analogy.
Where the evidence stops
Fifty people, then thirteen; correlational; nothing to do with AI. It supports a narrow concern, not a general claim of decline.
Used for
The narrow, once-used analogy that among fifty regular drivers, heavier lifetime GPS use was associated with worse spatial memory when navigating without it, and that of thirteen retested three years later, the heavier users declined more steeply — supporting only the narrow point that when a delegated capacity changes, the change may be specific to the ability no longer exercised.
Source 6 of 7Peer-reviewed conference paper. Complete conference text (23 pages) read for this note; identity confirmed against the official record (DOI 10.1145/3544548.3580919).
Novices who wrote code with an AI generator produced better code and showed no measurable loss when they later modified code by hand
What it supports
The sample (69 novices, ages 10–17); the authoring-then-manual-modification design; the higher authoring performance for the AI group; and the absence of a measurable difference on the manual modifying task.
Where the evidence stops
Introductory programming, a short study, immediate measures; it does not establish long-term independent capability.
Used for
The study of 69 novices aged ten to seventeen in which each task written with an AI code generator was followed by one the student had to modify by hand; the AI-assisted group produced better code on the authoring tasks and showed no measurable loss on the independent hand-modification task.
Source 7 of 7Preprint. Verified across the primary abstract, the arXiv text and the technical report.
In a supervised school trial, AI tutoring slightly edged human tutoring on the first unaided problem — with every AI message reviewed by an expert
What it supports
The two tutored arms; the expert review of every AI-drafted message; and the 66.2 per cent versus 60.7 per cent new-problem result.
Where the evidence stops
An exploratory preprint; both arms received tutoring; the AI arm was human-supervised, so it compares supervised AI tutoring with human tutoring alone rather than assistance against no assistance; the outcome is immediate and platform-specific, not long-term transfer.
Used for
The exploratory British-schools preprint in which both comparison groups received tutoring; in the LearnLM condition every AI-drafted message was reviewed and could be revised or replaced by an expert tutor before reaching the student; on the first unaided problem in the next study unit the estimated success rates were 66.2 per cent for the supervised LearnLM condition and 60.7 per cent for human tutoring alone.
Best available evidence
LearnLM Team (Google) & Eedi (2025), AI tutoring can safely and effectively support students: An exploratory RCT in UK classrooms, arXiv:2512.23633.
A note on this subject
This essay is about the difference between what a person and a tool produce together and what the person could do alone. None of the sources here measures an adult's independent capability across years of AI assistance; that study does not yet exist. The educational studies (mathematics, introductory programming, school tutoring) are the closest separate tests, and they are not about the adult knowledge-worker at the centre of the essay. The essay says so.
Update policy
This page records the source basis used for the essay. This note will be updated when a material source status changes. An update to this note does not silently rewrite the published essay — substantive corrections will be disclosed.